/peachdb

Managed solution to calculate, store and use embeddings, with support for open-source and multimodal models.

Primary LanguagePythonApache License 2.0Apache-2.0

PeachDB

PeachDB - the AI-First, Embeddings Database

Managed solution to calculate, store and use embeddings,
with support for open-source and multimodal models.


Our core API has 4 functions

from peachdb import PeachDB

# Create a new PeachDB instance or reference an existing one
db = PeachDB(
    project_name="my_app",
    embedding_generator="imagebind",  # or "sentence_transformer_L12"
    embedding_backend="exact_cpu",  # "exact_gpu", or "approx"
    distance_metric="cosine",  # or "l2"
)

# Auto-compute & upsert embeddings at scale using the specified `embedding_generator` model
db.upsert_text(  # or "upsert_image" or "upsert_audio"
    csv_path="/path/to/local/csv",  # or "s3://path/to/csv"
    column_to_embed="foo",  # column values can either be string or public URI to image/audio
    id_column_name="id",
    embeddings_output_s3_bucket_uri=None,  # required when using S3 URI for `csv_path`
    max_rows=None,  # or N to process N rows
)

# Query top 5 similar results
ids, distances, results_df = db.query(
    query_input='An example query',  # or path to an image/audio file
    modality='text',  # "image" or "audio"
    top_k=5
)

# Deploy database as a publicly accessible FastAPI server
# GET /query?query_input='An example query'&modality=[text|image|audio]&top_k=5 to fetch 5 most similar results
db.deploy()

Why another vector embedding solution?

We've streamlined the entire end-to-end process of creating, storing, and retrieving embeddings, making it developer-friendly, seamless, and cost-effective. You no longer have to build custom pipelines or fret over hardware setups & scalability issues. Get started within minutes, and leave the worries of management & scale to us.

Key features include:

  • Automated, cost-effective & large-scale embedding computation: We can process millions of data points via scaling to 10s of GPUs easily. We do so via serverless GPU functions, which keeps costs low.
  • Multimodality: Non-text embeddings have made significant progress, but they're always harder to work with. PeachDB provides native support for image/audio/text from day one.
  • Open-source models: We abstract away the setup requirements for open-source models. In our experience, getting started with almost all open-source models typically requires a few hours. Our prebuilt containers solve this.
  • Highly Customizable: Tailor the DB to suit your needs. You can customize:
    • Embedding models: as described above.
      • ⏳ Coming soon: Multi-threaded support for OpenAI APIs
      • ⏳ Coming soon: Custom embedding functions for even more flexibility.
      • ⏳ Coming soon: Bring your own embeddings.
    • Backend: choose between exact_cpu (numpy), exact_gpu (torch), or approx (HNSW).
    • Distance metrics: cosine or l2.
  • Effortless Deployment: Deploy PeachDB as a publicly available server with a single API. No need to worry about nginx or SSL certs.
    • ⏳ Coming soon: Managed, scalable deployment.
  • Consistent API: Experience the same API across all environments - dev, test, and prod.
  • Open Source: Apache 2.0.

Example

Below is a walkthrough for creating a web server for a music recommendation app. To power the app, we are using the Kaggle 5M song lyric dataset

  • Ssh into your remote instance (doesn't need GPU)

  • Create & activate a new conda environment conda create -n spoti_vibe python=3.10 & conda activate spoti_vibe

  • Install PeachDB: pip install peachdb

  • Setup Modal

    • Create an account at modal.com
    • Install the modal-client package: pip install modal-client
    • Setup token: modal token new

  • (optional: for AWS S3) PeachDB accepts local & S3 paths to datasets for embedding computation. To use S3 URIs, ensure you've installed the aws cli and run aws configure. The credentials should have read & write access to the relevant bucket you plan to use.

  • mkdir spoti_vibe/

  • Create a new module server.py inside the directory

  • Add the following code

    from peachdb import PeachDB

import os

# Fetch the username & key by creating a new API token at https://www.kaggle.com/settings
os.environ["KAGGLE_USERNAME"] = None  # set user name
os.environ["KAGGLE_KEY"] = None  # set key

import kaggle  # make sure you've run `pip install kaggle`

kaggle.api.authenticate()
# It can take a few mins to download depending on the network speed
kaggle.api.dataset_download_files("nikhilnayak123/5-million-song-lyrics-dataset", path=".", unzip=True)

db = PeachDB(
    project_name="spoti_vibe",
    distance_metric="cosine",
    embedding_backend="exact_cpu",
    embedding_generator="sentence_transformer_L12",
)
db.upsert_text(
    csv_path="./ds2.csv",  # dataset name as observed on Kaggle
    column_to_embed="lyrics",
    id_column_name="id",
)

db.deploy()  # Public URL will be printed to console

And that's it! You should now have a publicly available server that can listen to query requests from the user on:
GET <PUBLIC_URL>/query?text='Happy, upbeat summer'&top_k=5

Use-cases

FAQs

Q) How can I delete a project? Run db.delete(project_name="my_app")

Get Involved

We welcome PR contributors and ideas for how to improve the project.

Special Thanks

To Modal, DuckDB & pyngrok for developing wonderful services

License

Apache 2.0